Energy PASS, tension near-pass, valence FAIL on the corrected instrument -
kept as the warm-start seed for the v1.1 story-driven continuation.
Co-Authored-By: Ziggy <ziggy@artificialhumanity.io>
Claude-Session: https://claude.ai/code/session_01RUFyXDHbjtt9ong7EWZAqy
derisk-energy-24k: replace pickle checkpoint with safetensors weights
lmcfarlin
2dc3f3a
HF picklescan flags Lightning ckpts (functools.partial in callback state).
Weights extracted bit-exact (338 tensors); hparams recorded as JSON; the
resumable original stays on the training host.
derisk-energy-24k README: resolve stale open-questions section
lmcfarlin
8fcf838
Registry: mark both 24k promotions as published (owner call 2026-07-16)
lmcfarlin
5d1b2d4
derisk-energy-24k: LiteRT split-graph lane (spk + vat inputs, 24k vocoder)
lmcfarlin
a38cd60
toolchain: /data/toolchain/litert-conversion/convert_vat.py (extends the
v1-ljspeech build_matcha recipe). New graph inputs: spk(1,64) host-lookup
vector (spk_emb.bin, 247x64) + per-token/per-frame vat(1,3,T); decoder t_emb
is 224-dim (multi-speaker widens the U-Net input). Vocoder graph is the
fine-tuned 24k HiFi-GAN g_02510000.
Gates: wrapper-vs-model max|diff| <=1.7e-6/0.0 (FiLM paths included);
per-graph fp16 corr 1.000000/1.000000/0.999999; e2e waveform corr
0.9993-0.9999 (3 speakers x 3 energies); energy monotonic through the full
TFLite pipeline; all graphs GPU-clean (repeat()->broadcast-MUL rewrite
removed the only BROADCAST_TO blockers).
Promote derisk_energy checkpoint_epoch=099 (staged, not pushed)
lmcfarlin
b36c69d
train-repo: Artificial-Humanity/Sonora, MLflow run nimble-duck-437
gate verdict in derisk_energy_gate: all 4 sweep groups pass north-star §7
thresholds (energy rho ~1.000, WER delta <= 0.042, leakage <= 0.091) with
real margin. Sweep renders human-audited 2026-07-16, energy/loudness change
judged clearly discernible and natural.
Staged locally only -- same open question as vocoder-24k-hifigan/: this is
ONE VAT channel (energy), not the full directable actor, so whether a
partial-VAT checkpoint belongs in the public HF repo is a deliberate call,
not a default push. No auto-queued follow-on training this time either --
next steps are scope decisions (full-VAT corpus question, export-harness
adaptation), not mechanical continuations.
Promote HiFi-GAN 24k/80-band vocoder checkpoint g_02510000 (staged, not pushed)
lmcfarlin
2a30eaa
train-repo: Artificial-Humanity/Sonora @ a34dcea
run: MLflow hifigan_24k (vocoder_training), gate verdict in hifigan_24k_gate
Convergence: WER 0.073 <= floor+margin 0.084, mel-L1 delta 0.81% < 5% plateau.
A/B pairs human-audited, judged indistinguishable from ground truth.
Staged locally only -- see README for the open question on whether a
standalone (unintegrated) vocoder checkpoint should go to the public HF repo
before it's paired with a 24k acoustic model into a deployable e2e graph.
docs: lowercase Reference/models in consumption steps (workspace convention)
Lloyd McFarlin
d8ce820
docs: refresh model card (provenance, consumption) + ship engine config.json
Lloyd McFarlin
ff5898a
- New Provenance section: adopt the per-promotion convention (commit
messages record train-repo SHA + MLflow run) and backfill v1-ljspeech
best-effort (run lyrical-goat-689, export-fix a537e03).
- Fix stale consumption steps: the workspace file is now sonora.tflite
under Reference/Models/ resolved via prosodia_models.json roles;
drop the bogus cargo invocation for apps/tuner/build.sh.
- Add v1-ljspeech/config.json (locked 178-symbol vocab, sample_rate
22050) so the e2e artifacts are self-describing — the Prosodia engine
reads this adjacent to the model file. bootstrap.sh now fetches it.
- YAML metrics: add wer (ASR WER 0.000 on the fixed float32 export).
train-repo: n/a (docs/config only) run: n/a
feat: add LiteRT split-graph lane (litert-torch, mobile/delegate-ready)
TF post-training float16 quantization of the corrected graph (59.4 MB vs
187 MB f32), all I/O tensors f32, tensor names restored to the engine
contract via Sonora scripts/rename_tflite_tensors.py. Verified: deterministic
ONNX-parity cosine 0.9991, stochastic ASR parity with the ONNX oracle.
README: recommend this for mobile/edge; demote the f16-I/O onnx2tf build to
delegate experiments.
Co-Authored-By: Ziggy <ziggy@artificialhumanity.io>
Claude-Session: https://claude.ai/code/session_01BxJbryMycuFStzRrJULs1W
fix: re-export TFLite artifacts with corrected encoder LayerNorm
